Kai Fukami (UCLA) -- Flow field reconstruction from sparse sensors with machine learning

Global field reconstruction from sparse measurements is a long-standing challenge in fluid dynamics. The information from the reconstructed flow fields can be leveraged not only for understanding but also for controlling fluid flows. While conventional linear-theory-based techniques have played a crucial role, neural networks have recently been utilized to achieve a robust reconstruction. Neural networks can efficiently consider strong nonlinearities and multi-scale nature in reconstructing fluid flows. For example, we have performed neural-network-based super-resolution analysis, which reconstructs a high-resolution flow field from its low-resolution counterpart, and have demonstrated with turbulent flows. This talk will introduce our recent studies in extending super-resolution reconstruction of fluid flows for (1) establishing a robust model against on-/offline sensor situations assisted with Voronoi tessellation, (2) performing a qualitative reconstruction with small training data sets with transfer learning, and (3) quantifying uncertainties of machine-learning-based estimation with mixture density network.

Our methods will be demonstrated with a wide range of fluid flow examples from laminar to turbulence.